The introduction of networks into technical systems for facilitating remote data transmission, low complexity in wiring and easy diagnosis and maintenance, raises new challenges in fault detection (FD), such as how to handle network-induced time-varying transmission delays, packet dropouts, quantization errors and bit errors. These factors lead to increasing interest in developing new structures and design schemes for FD of technical systems over networks. In this thesis all network-induced effects are analyzed and modeled systematically at first. By observing the stochastic inheritance of networks, an FD framework of Markov jumping linear systems is presented as a basis for the later developments. Then two observer-based schemes for the purpose of FD over networks with guaranteed false alarm rate (FAR) are proposed: a remote FD system and an FD system of networked control systems (NCSs). The remote FD scheme is for detecting faults in technical systems at a remote site, where system measurements are transmitted via networks. In this scheme, the coding mechanism of communication channels is investigated from the view point of control engineering and new methods are developed for optimal residual generation and evaluation by considering network-induced data loss and corruption. A novel design scheme of FD system is also developed for NCSs, where the technical system is networked, i.e. controllers, actuators and sensors are connected with communication channels. In this scheme, network-induced transmission delays, packet dropouts, quantization errors are taken into account for the design of the optimal FD system. The linear matrix inequalities (LMIs) and convex optimization techniques are applied for assisting the design procedures. The developed schemes are tested with numerical examples and implemented in a three-tank system benchmark, and their superiority to existing solutions is demonstrated. Existing restrictions are overcome and new observer-based FD schemes over networks are introduced having the following characteristics: (1) the residual generators in both schemes are optimal in the sense of achieving the best trade-off between sensitivity to system faults and robustness against system disturbances and network-induced effects; (2) the proposed schemes can provide reliability information of rising fault alarms by analyzing the mean and variance of residual signals. Such information is very useful for practical applications in industries; (3) the design of residual generators and computation of thresholds can be efficiently solved by means of existing LMI-solvers.